Decolonizing Linguistic Policies in Automated Speech Recognition: A Framework for Cross-Culturally Competent Speech AI

The paper identifies systemic biases in Automatic Speech Recognition (ASR) systems, particularly concerning low-resource, Indigenous, and non-standard language varieties. These biases are framed not merely as technical failures but as implicit linguistic policies perpetuating colonial language hierarchies. The authors introduce a 'Three Harms (3M) taxonomy' (Misrecognition, Misalignment, and Mistrust) and a seven-layer situatedness model to address linguistic diversity in ASR. The analysis indicates that current ASR frameworks, by determining whose voices are machine-legible, inadvertently exclude significant linguistic communities.

Why it matters

This research is strategically important because it exposes how technological design choices in ASR systems can perpetuate social inequities and hinder access to essential services for marginalized linguistic communities. Addressing these biases is crucial for fostering inclusive digital infrastructure and ensuring that technological advancements benefit all segments of society, preventing the exacerbation of existing disparities.

Key insights

• ASR failures for diverse language varieties are attributed to implicit linguistic policies that reinforce colonial language hierarchies, rather than solely technical errors.
• The paper introduces a 'Three Harms (3M) taxonomy' to categorize the negative impacts of biased ASR systems: Misrecognition, Misalignment, and Mistrust.
• Data, metrics, and model priors in ASR design inherently determine which voices achieve machine legibility, leading to exclusion of certain linguistic groups.
• The research proposes a seven-layer situatedness model as a framework for incorporating linguistic diversity into ASR and ASR-mediated voice interfaces.
• ASR systems influence access to critical public services, healthcare, and education, making their linguistic biases a significant societal concern.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.06141